Glaze tool lets artists cloak work against AI style-mimicry

A team at the University of Chicago led by computer science professor Ben Zhao released Glaze, a free tool that lets artists add near-invisible alterations to their artwork before posting it online, designed to disrupt an AI model’s ability to learn and reproduce that artist’s style. The underlying research paper, “Protecting Artists from Style Mimicry by Text-to-Image Models,” had been posted to arXiv the previous month and was later presented at the USENIX Security Symposium, where it won a Distinguished Paper Award and co-won the 2023 Internet Defense Prize.

Text-to-image models could already be fine-tuned on a handful of an artist’s images to generate new work convincingly imitating their distinctive style, without permission or compensation — a use case that had provoked protest from illustrators since Stable Diffusion’s release the previous year. Glaze worked by computing how a given AI model perceived the artwork’s style internally, then applying a “cloak” of pixel-level perturbations, invisible to a human viewer, that shifted those internal features toward a different target style. A model trained on cloaked images would learn the wrong stylistic signature, degrading its ability to mimic the artist convincingly, while a person browsing the same image online would see no visible difference.

Glaze offered artists a defensive option in a fight they otherwise had almost no technical leverage over, at a moment when opt-out mechanisms and legal remedies against style mimicry were unsettled or nonexistent. The same University of Chicago group followed it in January 2024 with Nightshade, a complementary tool designed to actively corrupt a model that trained on protected images rather than merely evade its style-copying.